About Saibro
Saibro plans the soil sampling rounds of enhanced rock weathering projects.
About Saibro
Its user is the MRV lead who must turn lab sheets into a claim. The protocol most such projects follow accepts only measured removal, sizes each Unit by power analysis and credits at the 30th percentile, so a first round too noisy to call means a second at several times the density.
Saibro's planner reads the baseline fingerprint, the rock analysis, the spreading log, each round's lab sheets, and public rainfall and soil maps. It forecasts weathering per field with an interval and proposes how many cores each zone needs for its Unit to reach 80% power; a seeded random draw places them. With cores back, it computes the claim by the protocol's arithmetic.
Rock weathers unevenly across one field, and predicting where is the job Saibro gives to machine learning. A network pretrained on published field trials and public soil and climate data, refitted on each project's cores, maps expected signal and its spread zone by zone; the spread decides where cores go. A smaller document model reads each lab's sheets. Power analysis and the claim stay in plain code.
No Saibro forecast becomes a claimed tonne, and no plan or claim leaves without the MRV lead's approval. The model never drops a field or removes an outlier on its own, and each period it must place nine in ten cores from held-out fields inside its 90% range, or plans revert to plain power analysis. Each round's pairs of forecast and measured signal refit that project's model, and train our shared one only if the developer opts in.
Saibro's computing follows the lab calendar, so it comes in bursts. A round's sheets landing means every Unit refitted and thousands of candidate plans simulated; a period's end, 500 rarefaction repeats and 10,000 Monte Carlo draws per Unit; a new project, ten years of land-use history. Between rounds little runs. Parcels and farmer names, personal data under the GDPR, go to no outside model provider: inference and training both happen on cloud GPU capacity we control in N. Virginia.
Saibro rents cloud GPU capacity in N. Virginia by the hour through February 2027, while the first harvest-season rounds show how large a burst gets. In March 2027 the shared network retrains on opted-in pairs; from July 2027, before the autumn rounds that follow the maize harvest, a year-round GPU baseline in N. Virginia is reserved, with hourly capacity added for period ends.
The founders
Leonor Caeiro, Duarte Pombinho and Filipa Valadas founded Saibro in Évora in April 2026. The planner and the Android field app have been in beta since August 2026 and the claim page since October; Saibro is taking developers whose next sampling round is still ahead.
Leonor Caeiro
President
Ran soil sampling crews for a soil carbon project developer on Alentejo farms, and spent the season after its first rock spreading trial re-coring fields whose lab results could not tell weathering from noise.
Duarte Pombinho
CAIO
Wrote his doctorate in Évora on mapping soil properties from sparse cores, then built yield models for a group of maize growers in the Ribatejo before turning the same maps on weathering.
Filipa Valadas
CDO
Kept the results database of a Lisbon soil lab for eight years, where every client's sheets came in a different layout and sample labels went missing between the field and the bench.